Reproduce “Advancing Spatio-Temporal Processing in Spiking Neural Networks through Adaptation” on OpenResearch

About this paper

Maximilian Baronig, Romain Ferrand, Silvester Sabathiel, Robert Legenstein

This research introduces the Symplectic Euler discretization for adaptive Leaky Integrate-and-Fire (adLIF) neurons, ensuring stable and expressive spatio-temporal processing in spiking neural networks. The approach achieves state-of-the-art performance on various tasks while demonstrating inherent robustness and simplifying neuromorphic hardware deployment by negating the need for explicit normalization.

Key insights

• The Symplectic Euler discretization ensures asymptotic stability for adLIF neurons across their full frequency range and parameter space, resolving critical issues of the common Euler-Forward method.
• adLIF neurons possess an intrinsic inductive bias towards detecting high-fidelity temporal changes and precise feature timings in input spike trains, making them highly effective for spatio-temporal processing.
• The adaptation mechanism in adLIF neurons confers an inherent robustness and stability to network activity, often eliminating the need for explicit normalization layers during training and deployment.

Summary by alphaXiv.